#2: Objective vs. Search: Decomposing Tokenisers + Zeroth-Order Alignment
In this episode, Alex and Sam unpack why search procedure dominates over compression objective in tokenizer quality, and explore ComPO: zeroth-order LLM preference optimization without backpropagating through preference loss.
Papers Discussed
- • Objective vs. Search: Decomposing What Makes a Good Tokeniser (Ahmetcan Yavuz et al.)
- • A Zeroth-Order Paradigm for LLM Preference Alignment (Peter Chen et al.)